What an AI feature costs to run, per user, with October 2026 prices

Each time a user asks an AI product for an answer, the product sends text to a model, the model sends text back, and the provider charges for both. Across every user and every month, that becomes a cost line that grows with the business. This page shows how to estimate it before you set your own price. It is part of the guide on building your AI product after you raise.

How model providers charge

Providers charge per token. A token is a small piece of text, often a word or part of a word, and the provider's software splits every request into tokens before the model reads it. Prices are quoted per million tokens, with one price for input, the text you send, and a higher price for output, the text the model writes.

The table gives list prices for some models from the three large providers, in US dollars per million tokens, read on 26 October 2026.

Provider and model	Input	Output	Note
OpenAI gpt-6-astra	$10.00	$50.00	Up to 272K input tokens [1]
OpenAI gpt-6.1-sol	$2.00	$10.00	Up to 272K input tokens [1]
OpenAI gpt-6.1-terra	$1.00	$5.00	Up to 272K input tokens [1]
OpenAI gpt-6-luna	$0.10	$0.50	Up to 272K input tokens [1]
Anthropic Claude Opus 5.5	$4.00	$20.00	[4]
Anthropic Claude Sonnet 5.5	$2.00	$10.00	[4]
Anthropic Claude Haiku 5.5	$0.10	$0.50	Prompts up to 100,000 tokens; $0.50 and $2.50 above [4]
Google Gemini 4 Argon	$2.00	$10.00	Through November 30, 2026; $4.00 and $20.00 from December 1, 2026 [7]
Google Gemini 3.8 Flash	$0.75	$3.75	Through December 31, 2026; $1.50 and $7.50 from January 1, 2027 [7]
Google Gemini 3.5 Flash-Lite	$0.30	$2.50	[7]

OpenAI charges more above 272K input tokens in one request, which it calls long context [1]. Google's output prices include "thinking tokens", the text a model produces while working out its answer [7]. Google also says Gemini 4 Argon "is not available on the Free tier" [7].

Scheduled prices can also change. Anthropic's pricing page says that for Sonnet 5, an older model, "The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur" [4]. Read each price page on the day you decide.

A worked example

This example is invented. No source gives the token count of a real product, so replace these figures with your own logs as soon as you have them.

The product is a support assistant. Each request sends 3,000 input tokens: 2,000 tokens of fixed instructions and policy text that are the same every time, plus 1,000 tokens of the user's question and history. The model writes 500 output tokens. Each user sends 40 requests a month.

The arithmetic for gpt-6.1-sol is 3,000 divided by one million, times $2.00, which is $0.006 for input, plus 500 divided by one million, times $10.00, which is $0.005 for output: $0.011 per request. Forty requests make $0.44 per user per month, and 1,000 users make $440 a month.

Model	Cost per request	Per user per month	Per 1,000 users per month
Claude Opus 5.5	$0.022	$0.88	$880
Gemini 4 Argon from December 1, 2026	$0.022	$0.88	$880
gpt-6.1-sol, Claude Sonnet 5.5, Gemini 4 Argon to November 30, 2026	$0.011	$0.44	$440
gpt-6.1-terra	$0.0055	$0.22	$220
Gemini 3.8 Flash to December 31, 2026	$0.004125	$0.165	$165
Gemini 3.5 Flash-Lite	$0.00215	$0.086	$86
Claude Haiku 5.5, gpt-6-luna	$0.00055	$0.022	$22

The cheapest model in the table costs one fortieth of the most expensive. Whether the cheaper model is good enough is a question for your eval suite, a set of written, repeatable tests of the answers on your own cases. Run the same cases on two or three models and choose the cheapest one that passes. The guide on LLM evals explains how to build that suite.

The same text can produce a different number of tokens on different models. Anthropic's pricing page says Claude 4.7 and later models use a newer tokenizer, the software that splits text into tokens, which produces more tokens for the same text than earlier models, and that "The exact increase depends on the content and workload shape" [4]. Count tokens on the model you will use.

Prompt caching lowers the price of repeated text

In the example, 2,000 of the 3,000 input tokens are the same in every request. Prompt caching lets the provider store that repeated part and charge less for it when it is sent again. Each provider runs caching differently.

Provider	What its documentation says
OpenAI	Caching "is enabled by default for supported OpenAI models", with reused tokens "discounted up to 95%"; the minimum cacheable prompt is 1,024 tokens for GPT-5.6 and later [2]. Writing to the cache costs 1.25 times the normal input rate for the default 30 minutes, or 2.5 times for 12 hours [2].
Anthropic	A cache hit costs 10% of the standard input price, or 5% on Claude Opus 5.5 and Sonnet 5.5; writes cost 1.25 times for 5 minutes or 2 times for 1 hour; caching "pays off after one cache read for the 5-minute duration" [4]. The minimum is 512 tokens for its newest models, and shorter prompts are processed "without caching, and no error is returned" [5].
Google	"Implicit caching is enabled by default for all Gemini 2.5 and newer models", with a minimum of 4,096 input tokens on Gemini 4 Argon [8]. Explicit caching, where you create the cache yourself, is not supported for Gemini 4 Argon and later [8].

Apply it to the example. On Claude Sonnet 5.5 and on gpt-6.1-sol, a cached input token costs $0.10 per million [4] [1]. If the 2,000 fixed tokens are read from the cache, a request costs 2,000 tokens at $0.10, plus 1,000 tokens at $2.00, plus 500 output tokens at $10.00: $0.0072 instead of $0.011. For 1,000 users that is $288 a month instead of $440. The figure leaves out the cost of writing the cache, which is paid when the cache is created and again each time it has to be written after it expires.

The minimums decide what is cached. In this example the fixed block of 2,000 tokens is above OpenAI's 1,024 and Anthropic's 512, and below Google's 4,096 for Gemini 4 Argon, so on Argon this block would not be cached. Anthropic notes one more effect: for most of its models, only uncached input tokens and cache creation tokens count toward the limit on input tokens per minute, which makes caching what it calls "an effective way to increase your effective throughput" [11]. Throughput is how many requests the product can send in a given time.

Batch requests cost half

Some work does not need an answer while the user waits: a nightly summary, a report, a large set of documents to sort. Each provider sells a batch service for that work at half price. OpenAI describes a "50% cost discount compared to synchronous APIs" and says "Each batch completes within 24 hours (and often more quickly)" [3]. Anthropic charges "50 percent of the standard API prices", says most batches complete "within 1 hour", limits a batch to 100,000 requests or 256 MB, and says batches "expire if processing does not complete within 24 hours" [6]. Google lists "Batch API (50% cost reduction)" on its pricing page [7] and says "The target turnaround time is 24 hours" [9].

In the example, gpt-6.1-sol in batch costs $1.00 for input and $5.00 for output [1], so a request costs $0.0055 and 1,000 users cost $220 a month. Anthropic adds that "Batch API and prompt caching discounts can be combined" [4].

Charges that add to the price

Three options raise the price.

• Keeping data in a region. OpenAI says its "Regional processing (data residency) endpoints are charged a 10% uplift for models released on or after March 5, 2026" [1]. Anthropic says US-only inference on Claude 4.6 and later models "incurs a 1.1x multiplier on all token pricing categories" [4]. In the example, if the model is covered, a 10% uplift takes $440 to $484 a month for 1,000 users. An uplift is an added charge.
• Faster answers. OpenAI's Fast mode, which it says was renamed from Priority processing on July 30, 2026, lists gpt-6.1-sol at $4.00 for input and $20.00 for output [1], twice the standard price for that model. The example would cost $880 a month.
• Promotions that end. OpenAI says "GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026" [1]. A price set during a promotion can rise when it ends.

Spending limits at each provider

Each provider sets limits on a new account that rise as you spend more. Google says its spend-based limits protect against unexpected charges, and any of these limits can also stop your product when use grows faster than the limit.

Provider	Tier	What it takes or allows
OpenAI	Build	Reached after $5 in total credit purchases; $500 a month usage limit [10]
OpenAI	Grow	Reached after $500 in total credit purchases; $200,000 a month [10]
Anthropic	Start / Build / Scale	Monthly spend limits of $500, $1,000 and $200,000 [11]
Google	Tier 1 / Tier 2 / Tier 3	Spend-based limits of $10, $50 and $200 per 10 minutes [12]

Check the tier before a launch. OpenAI and Anthropic both let you set your own lower spending limit inside the tier [10] [11]. Anthropic's rate limits page says that once you reach your tier's spend cap, "API usage pauses until 00:00 UTC on the first day of the next month, unless you request a higher limit sooner" [11].

From cost per user to gross margin

Gross margin is the share of revenue left after the cost of serving the product. Suppose, as an invented example, you charge $20 per user per month. At $0.44 of model fees per user, the model costs 2.2% of revenue. At $0.88 it costs 4.4%. Hosting, storage and support add to the cost of serving, and heavy users can send many more than 40 requests, so measure the cost for your heaviest 10% of users as well as the average. Investors rebuild this margin from invoices, as the page rebuilding inference gross margin from invoices explains.

Reveneau's view

Reveneau's view is that a founder should know the model cost per user before setting a price, and should measure it again every month from real logs. Log the input and output tokens of every request from the first day. Choose the cheapest model that passes your eval suite, put fixed text first in each request so it can be cached, move work that can wait to batch, and set a spending limit at each provider. The page API or your own model for a startup covers the choice of model, and engineering budget after a raise shows where the model line sits in the budget.

Reveneau is an AI software development consultancy with offices in San Francisco and Vancouver. Reveneau says AI writes all of its code and every change must pass an eval suite written from the specification, graded by Jev, TypeSafe AI's decision model, which Reveneau says runs the suite ten times faster than its previous grader. Reveneau says it uses AI instead of hiring more engineers and puts that saving into clients' prices. To plan an AI feature with its running cost measured, see AI development at Reveneau or contact Reveneau.